Patchwork: A compact representation for 3D polygonal shapes

Patchwork is a general-purpose geometric representation built to model 2D and 3D shapes with a small parameter count. The paper says it comes with a rigorous mathematical framework, provable complexity bounds, and the ability to approximate arbitrary shapes in any dimension to arbitrary precision. That makes it more than just another neural shape encoding: the authors are explicitly trying to balance expressiveness, compactness, and theoretical guarantees.

For developers, the practical angle is the fitting pipeline. They propose a gradient-based optimization method plus a regularization loss that progressively removes redundant elements, so the representation becomes more compact after training instead of bloating with unused parts. The paper also calls out native inside-outside classification, which is handy for reconstruction, collision-style queries, and any workflow that needs a clear notion of solid vs. empty space. The implementation is reportedly available, and the work was...